I don’t expect that augmenting an AI with what it already knows and previously said will do much to improve its abilities.
The AI model has a limited context length - the amount of information that can be supplied along with your user’s question. Each API call is unique, and we give the illusion of memory by also supplying just a bit of previous chat. It is not enough to even maintain the start of a long ChatGPT conversation.
Therefore, if one wants to implement real data augmentation, where you are providing new knowledge to the AI, you’d typically process and chunk company knowledge, and then put it in an embeddings-based semantic search vector database. If someone asks when the Duluth location shipping dock is open, the AI can be injected with a search-like automated knowledge retrieval by finding what is similar to the question they asked.
People also imagine that ChatGPT is learning about their mental health conditions and becoming their friend, when in fact it might only see five of the last conversational turns. So you can do a bit of prompting to make the AI more sympathetic and seemingly knowledgeable if you want to play into human weakness.